Temporary Labeling for Learning Data Generation

The patent introduces a temporary label as an intermediary category between known classification labels and unclassified objects. When the learned model evaluates an object image and determines it does not match any known label (or matches with low reliability), the system assigns a temporary label instead of discarding the data. This intermediary label allows the system to maintain high classification accuracy for known objects while adapting to handle unknown objects, thereby resolving the contradiction between measurement precision and adaptability.

The patent segments the learning data collection process into two distinct pathways: one for high reliability objects (assigned known labels) and another for low reliability objects (assigned temporary labels). This segmentation allows the system to maintain strict quality control for confidently classified objects while simultaneously expanding the quantity of learning data by including uncertain objects with temporary labels. The segmented approach resolves the contradiction by allowing both high-quality and quantity-expanding data collection strategies to coexist.

The patent implements a self-service labeling system where the learned model automatically evaluates object images and assigns labels (either known labels or temporary labels) without requiring manual intervention for every image. The system serves itself by using its own learned knowledge to generate labels, thereby maintaining high labeling accuracy while dramatically reducing the time and computational resources required compared to manual labeling of all images.

The patent extracts and separates temporary label data from the main labeled dataset, treating them as distinct but complementary components. By extracting temporary labeled data as a separate category, the system can maintain the homogeneity of the primary labeled dataset while still incorporating the quantity benefits of temporary labels. This extraction approach allows the system to manage data heterogeneity by clearly delineating between confidently labeled and temporarily labeled data, resolving the contradiction between quantity increase and homogeneity maintenance.